基于特征统计可分性的遥感数据专题分类尺度效应分析

现有的对地观测遥感卫星能够提供从0.61 m到数十公里空间分辨率的遥感数据。通过遥感数据的专题分类得到的专题图的精度不但受遥感数据光谱特征、遥感数据处理和分类过程的影响, 而且受到所用的遥感数据的空间分辨率的影响。遥感数据空间分辨率的变化对遥感专题分类精度的影响受混合像元数目的变化和类内光谱特征变异程度的变化这两个矛盾的因子影响。空间分辨率对分类精度的最终影响决定于这两个矛盾影响因子的净效应。通过分析遥感专题分类中分类特种的统计可分性随遥感数据空间分辨率的变化来分析空间分辨率变化对分类精度的净效应。采用变换的离散度作为特征的统计可分性度量。以TM数据进行土地利用/土地覆被分类为例,首先将原始分辨率的图像以简单平均方法逐步尺度扩展到不同分辨率,然后在原始空间分辨率的图像上,根据该地区土地利用图进行层次随机采样,并以原始分辨率图像上的随机采样位置为掩模,在尺度扩展后的图像上进行同样位置的随机采样,最后在各空间分辨率上分别计算类对间的变换离散度。对变换的离散度随空间分辨率变化的规律进行了分析和定性解释。研究表明,类对间空间邻接结构对类别间混合像元数目随空间分辨率的变化有决定性影响;不同类对之间的最大统计可分性可能发生在不同的空间分辨率;空间分辨率越高,并不一定分类精度越高;不同类别之间的分类需要不同空间分辨率的数据。

Exploring the Scale Effect in Thematic Classification of Remotely Sensed Data:the Statistical Separability-based Method

The statistical separability is used to explore the scale effect of remote sensing data classification and to determine optimal resolution in this paper. The Landsat TM image with 30m spatial resolution is up-scaled to different spatial resolutions. The stratified random sampling method was used to select the training samples at 30m resolution, and the location of training samples were saved as masks to take training samples for up-scaled images so that training samples for images at every resolution are at same location. The transformed divergence and J-M distance of training samples at every resolution were calculated for every class pair, and were plotted versus the spatial resolution. The landscape metrics of the land cover in the study area were calculated Analysis to these plots showed that, for different pair of classes, the change pattern of statistical separability with spatial resolution is quite different. The spatial pattern between pair of classes has significant effect on the statistical separability pattern of change with spatial resolution and can be used to explain the underlying reasons for the change patterns. For our experimental data, the average statistical separability reached to the maximum at the 60m spatial resolution, which means that finer spatial resolution not necessary lead to high separability.